Hi guys, so via the API, we ask GPT to create a piece of content in Italian. What we see is that it creates in english and then translates to Italian. So we get charged twice the amount of words…first the english version and then Italian. Anyone else face this?
How exactly are you observing this? Is the model outputting the same text back in English before it outputs the Italian version?
Or are you seeing the cost higher than you expected in the API Usage page?
Yes we are seeing the cost higher than expected in the API usage. I am wondering if we are doing something wrong here or is this normal?
We are for sure seeing more tokens being used for other languages than English.
This seems like a prompt issue, I’m pretty sure that can be fixed
Yeah - this is somewhat expected. The tokenizer is optimized for english text, probably since it’s the most common language. If you take a story and translate it into multiple languages and then count the tokens, you’ll see the massive disparity:
Are you aware that the API charges for both the tokens in and tokens out?
Providing an example of your prompt and the model response might help us understand.